Microscopic Image Dataset for CAT6 Cable Outer-Sheath Surface Condition Classification
Description
This dataset contains 2,400 original microscopic images acquired from the outer sheath of an approximately 10-m-long CAT6 Ethernet cable under controlled laboratory conditions. The cable had been in use for an extended period and exhibited different surface conditions along its length. Images were captured using a USB digital microscope with a 1920 × 1080 pixel image sensor under consistent illumination, working distance, and viewing conditions. The dataset is organized into three balanced surface-condition classes: Intact, Bending_Damage, and Severe_Damage, with 800 images in each class. Intact images show no visible surface damage; Bending_Damage images exhibit localized texture distortion or surface waviness without visible material rupture; and Severe_Damage images contain visible cracks, material rupture, or deep surface discontinuities. All 2,400 images are original microscope acquisitions. No images were generated through data augmentation, extracted from video frames, or produced by cropping larger source images. The images were acquired across multiple physical regions of the same cable; however, physical region identifiers were not recorded during acquisition. The dataset is intended as a task-specific resource for supervised microscopic classification of CAT6 cable outer-sheath surface conditions and for comparative evaluation of image-classification methods under controlled acquisition conditions. Because the dataset originates from a single cable specimen, it should not be interpreted as representing cross-cable or real-world generalization. Future extensions may incorporate multiple cable specimens, materials, manufacturers, acquisition conditions, and region-level metadata. The original 2,400 images are provided in the Raw_Dataset.zip archive, organized into three class-specific folders: Intact, Bending_Damage, and Severe_Damage. Citation: If you use this dataset in your research, please cite the associated research article. The full citation and DOI of the article will be added to this dataset record upon publication.
Files
Steps to reproduce
The dataset contains 2,400 original microscopic images organized into three class-specific folders: Intact, Bending_Damage, and Severe_Damage, with 800 images per class. For the reported benchmark experiments, the images were partitioned using stratified random sampling into 70% training, 15% validation, and 15% test subsets, using a fixed random seed (rng(42)). This resulted in 1,680 training, 360 validation, and 360 test images. Images were resized to 224 × 224 pixels before being provided to the CNN models. No data augmentation was applied. Five ImageNet-pretrained CNN architectures—MobileNetV2, ResNet18, GoogLeNet, DenseNet201, and EfficientNetB0—were evaluated using a common training protocol with SGDM, an initial learning rate of 0.0001, a mini-batch size of 16, and a maximum of 10 epochs. The training data were shuffled at every epoch, and validation was performed every 20 iterations. The exact image-level training, validation, and test assignments used in the experiments are provided in the accompanying CAT6_Exact_Train_Validation_Test_Splits.csv file. The CAT6_Split_Lists_README.txt file explains the split structure and file paths.
Institutions
- Bursa Uludağ Üni̇versi̇tesi̇Bursa Province, Bursa